인공지능의 학술적 글쓰기 성능 검토: 학술지 논문 초록을 중심으로

Abstract

This study examines the linguistic differences between humans and generative AI models in the specific genre of academic abstracts. Generative models, more prevalent in higher education, would help undergraduates and graduates develop their scholarly writing skills by offering lexical alternatives or structural suggestions, but the linguistic features of the texts produced by AIs, unlike issues such as hallucination and general bias, have not attracted sufficient academic attention. This study compared 212 English abstracts published in The Journal of Eighteenth-Century English Literature with the ones produced under identical prompts by three leading models (GPT-5, Claude Sonnet 4.5, and Gemini 2.5 Pro). The results reveal that AI-generated abstracts, which tend to be longer and verbose than those written by humans, display inflated proportions of nouns and adjectives in contrast to the lower rates of pronouns and modal verbs. N-gram analysis further shows a heavy reliance on recurrent phrases such as “close reading” or “examine,” implying a narrow stylistic range. These findings reveal that the three models cannot write concise and easy-to-read abstracts because they do not only consider the purpose of the abstract writing but also employ rhetorical conventions properly. This limitation of the AI indicates that the educational potential of generative models deserve further academic investigation, but their use in class should remain experimental, pointing to the necessity of cultivating students’ critical literacy regarding both the strengths and limitations of AI-assisted writing.